Paul Graham says Replit invented vibe coding in 2024; Masad says Replit first ran LLMs to completion.
Replit's reversible file system and multiplayer design let an agent act as another user, winning enterprise deals.
Masad says AI commoditizes, forcing reinvention, and predicts a jump from one agent per person to 10,000 agents.
Replit Bets AI Coding Agents Should Run to Completion
Replit bet that AI coding agents should run to completion rather than one turn, automating a loop its users were already performing by hand. Its kernel and file-system work made the agent reliable enough for enterprises, where it turns $10m projects into $250,000 ones. Masad argues AI commoditizes everything, so reinvention is mandatory and firms will reorganize into self-driving companies. Graham adds that easier programming produces superlinearly more programmers, and the best startups reinvent themselves to get 100 times bigger.
Takeaways
Replit rebuilds $10m projects for $250k. Big banks and multinationals pay $10m for SaaS or old software, Masad says. Replit's agent and infrastructure let it do projects scoped at that price for $250,000. That economics, he argues, kills the outsourcing industry and made Replit Bounties redundant.
Systems depth wins enterprise AI coding deals. Replit employs kernel and operating system engineers and built a transactional, reversible file system. Multiplayer collaboration meant an agent could simply be another user. Enterprises care about reliability and security, so that depth matters, even if Replit's guardrails make products slower and more expensive.
Rationing AI tokens per person is wrong. Masad disagrees with capping spend at $100 a month per employee and says companies should measure return instead, because one person can drive 100 times the value. He predicts going from one agent per person to 10,000 agents, which cannot run on a MacBook.
AI commoditizes, so reinvention is mandatory. Masad says everything Replit does in AI is constantly commoditized, so it must be doing the next thing. Replit stopped describing itself as vibe coding and published The Self Driving Company, a vision for how firms reorganize as AI spreads.
Programmers face denial, grief, then acceptance. Masad says he grieved the craft of programming in 2020. Engineers will go through denial and depression, then acceptance. Graham found the demo aesthetically offensive and recalls being told not to look at the code because it was object code.
Easier programming yields superlinearly more programmers. Graham argues the relationship between how easy programming is and how many programmers exist is superlinear. The best startups reinvent themselves and get 100 times bigger, he says, like surfing a tidal wave rather than clinging to a palm tree.
How Replit built the agent
Users were already generating 50-60 percent of their code with autocomplete, and some did 100 percent of their coding with AI before the agent existed.
Masad invited users to the office and watched them copy code into Replit, copy errors into ChatGPT, and copy fixes back, a manual loop.
Replit automated that loop with the running program as the terminal condition, unlike ChatGPT and most tools that did one turn.
The 2024 harness was ten times more complex than today's because the models were weaker, Masad says.
Replit launched the agent in September 2024, and Andrej Karpathy called it a field AGI moment.
Tap a name to go there · tap a line to see why they are linked